Infrared-Visible Person Re-Identification via Multi-Modality Feature Fusion and Self-Distillation

Most existing cross-modality person re-identification methods mine modality-invariant features, while ignoring the discriminative features inherent to each modality. To fully utilize the inherent features in different modalities, an infrared-visible person re-identification method via multi-modality feature fusion and self-distillation is proposed. Firstly, an attention fusion mechanism based on a dual classifier is proposed. This mechanism assigns greater fusion weights to the self-owned features of each modality, while conversely assigning lesser weights to the common features. This approach aims to obtain multi-modality fusion features that encapsulate the discriminative self-owned features of each modality. To enhance the robustness of network feature in adjusting to changes of pedestrian appearance, a memory storage is constructed to store the multi-view features of pedestrians. A parameter-free dynamic guidance strategy for self-distillation is also designed. This strategy aims to dynamically reinforce the multi-modality and multi-view reasoning capabilities of the network under the guidance of multi-modality fusion features and multi-view features. Finally, the network is able to infer the features of a pedestrian with different views of another modality from its single-modality image, thus improving the performance of the model for cross-modality person re-identification. Based on the PyTorch deep learning framework, comparative experiments are conducted with current main-stream methods on the public datasets SYSU-MM01 and RegDB. The results demonstrate that the proposed method achieves Rank-1 accuracies of 63.12% and 92.55%, respectively, along with mAP scores of 61.51% and 89.55%, respectively, which is superior to the comparison methods.